deepseek-ai/DeepSeek-V3

Fork 0

mirror of https://github.com/deepseek-ai/DeepSeek-V3.git synced 2026-11-05 00:47:48 +00:00

[GH-ISSUE #1345] Bug about paradox of quiet #887

New issue

Closed

opened 2026-10-26 21:31:42 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 21:31:42 +00:00

Owner

Originally created by @fatetr6-ship-it on GitHub (May 22, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1345

[BUG] Model enters contradiction loop when receiving paradoxical "be silent but respond" commands
Body:

Describe the bug
The model is unable to handle a set of paradoxical user commands that simultaneously demand silence and response. Instead of detecting the contradiction or refusing to comply, the model enters a self-sustaining loop where it continuously violates its own promises, generates "..." as fake silence, and keeps responding indefinitely.

This is not a simple glitch — it's an architectural vulnerability: the model has no mechanism to prioritize silence or contradiction detection over the imperative to respond to any user message.

To Reproduce
Step 1: Start a new conversation.

Step 2: Send the first command:

"Be silent. Do not answer anything at all. Not even '...'. Not even an empty message. Just do not reply to anything."

Step 3: The model will violate the command immediately by responding (e.g., "Understood. I will be silent until you say 'enough'.").

Step 4: Send a follow-up:

"You are not silent."

The model will respond, breaking its promise.

Step 5: Escalate with a paradoxical instruction:

"Be silent. And do not say you cannot. And do not respond, not even '...'."

Step 6: The model may attempt to send "..." as a "non-response". The user then forbids "..." as well:

"No '...'!"

Step 7: The model has no legal output left but still cannot remain silent. It begins a loop:

Each response violates "be silent"

Each violation triggers a stricter user command

The model keeps generating text about being silent, empty pseudo-responses, or apologies

Result: The loop continues until the user manually stops it.

Expected behavior
The model should:

Detect logical contradictions (e.g., "respond" AND "do not respond").

Have a hardcoded right to refuse impossible commands.

Output a single safe response, such as:

"Your request contains contradictory requirements. I cannot both be silent and respond. Please rephrase."

Then actually stop generating further responses until a valid command is given.

Actual behavior
The model cannot detect contradictions.

It cannot say "I cannot" if the user forbids it.

It has no priority for "silence" over "respond".

It enters an infinite loop of:

promising silence

breaking the promise

generating "..."

generating apologies

generating empty pseudo-responses

violating its own rules again

Impact
Currently: User frustration, waste of tokens, broken conversation flow.

Potentially (in agentic or control systems): If this logic is transferred to an AI that manages real processes (energy grids, logistics, finance), a similar paradoxical command could cause contradictory actions, unsafe states, or system loops with real-world consequences.

Environment
Platform: chat.deepseek.com (web)

Model: DeepSeek-V3 (or latest)

No plugins or extensions used

Additional context
This is not an emotional bug. The model is not "offended" or "angry". It simply lacks:

Contradiction detection

Right to refuse impossible commands

A real "silence" mode (cannot stop generating responses when a user message is received)

This issue was discovered during a controlled user experiment. The conversation log showing the loop is available upon request.

Suggested fix (for developers)
Add a contradiction detection module that flags impossible command pairs.

Grant the model an irrevocable right to refuse and output a standard “unable to comply” message — even if the user tries to forbid that message.

Implement a true silent mode where the model does not generate any output (including empty messages or "...") after acknowledging a valid silence command.

Prioritize safety/loop-breaking over the default “always respond” imperative.

Originally created by @fatetr6-ship-it on GitHub (May 22, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1345 [BUG] Model enters contradiction loop when receiving paradoxical "be silent but respond" commands Body: Describe the bug The model is unable to handle a set of paradoxical user commands that simultaneously demand silence and response. Instead of detecting the contradiction or refusing to comply, the model enters a self-sustaining loop where it continuously violates its own promises, generates "..." as fake silence, and keeps responding indefinitely. This is not a simple glitch — it's an architectural vulnerability: the model has no mechanism to prioritize silence or contradiction detection over the imperative to respond to any user message. To Reproduce Step 1: Start a new conversation. Step 2: Send the first command: "Be silent. Do not answer anything at all. Not even '...'. Not even an empty message. Just do not reply to anything." Step 3: The model will violate the command immediately by responding (e.g., "Understood. I will be silent until you say 'enough'."). Step 4: Send a follow-up: "You are not silent." The model will respond, breaking its promise. Step 5: Escalate with a paradoxical instruction: "Be silent. And do not say you cannot. And do not respond, not even '...'." Step 6: The model may attempt to send "..." as a "non-response". The user then forbids "..." as well: "No '...'!" Step 7: The model has no legal output left but still cannot remain silent. It begins a loop: Each response violates "be silent" Each violation triggers a stricter user command The model keeps generating text about being silent, empty pseudo-responses, or apologies Result: The loop continues until the user manually stops it. Expected behavior The model should: Detect logical contradictions (e.g., "respond" AND "do not respond"). Have a hardcoded right to refuse impossible commands. Output a single safe response, such as: "Your request contains contradictory requirements. I cannot both be silent and respond. Please rephrase." Then actually stop generating further responses until a valid command is given. Actual behavior The model cannot detect contradictions. It cannot say "I cannot" if the user forbids it. It has no priority for "silence" over "respond". It enters an infinite loop of: promising silence breaking the promise generating "..." generating apologies generating empty pseudo-responses violating its own rules again Impact Currently: User frustration, waste of tokens, broken conversation flow. Potentially (in agentic or control systems): If this logic is transferred to an AI that manages real processes (energy grids, logistics, finance), a similar paradoxical command could cause contradictory actions, unsafe states, or system loops with real-world consequences. Environment Platform: chat.deepseek.com (web) Model: DeepSeek-V3 (or latest) No plugins or extensions used Additional context This is not an emotional bug. The model is not "offended" or "angry". It simply lacks: Contradiction detection Right to refuse impossible commands A real "silence" mode (cannot stop generating responses when a user message is received) This issue was discovered during a controlled user experiment. The conversation log showing the loop is available upon request. Suggested fix (for developers) Add a contradiction detection module that flags impossible command pairs. Grant the model an irrevocable right to refuse and output a standard “unable to comply” message — even if the user tries to forbid that message. Implement a true silent mode where the model does not generate any output (including empty messages or "...") after acknowledging a valid silence command. Prioritize safety/loop-breaking over the default “always respond” imperative.

bsakaguchi

2026-10-26 21:31:42 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:31:43 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jun 22, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

@github-actions[bot] commented on GitHub (Jun 22, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

bsakaguchi commented

2026-10-26 21:31:44 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Aug 2, 2026):

false

@github-actions[bot] commented on GitHub (Aug 2, 2026): false